To be completely honest, we didn't do any profiling on this yet, so we don't really know if the AOSP build is spending a lot of time waiting for a VSYNC that doesn't arrive or something.
What we do know, however, is that a real GPU-bound 3D benchmark such as GLMark2 is not all that much faster (Not much of a surprise, given we don't have the source for the SGX driver, so we can't fix it -- it's about 2% faster), but other CPU and memory intensive benchmarks (e.g. Sunspider) show an almost equal improvement on our builds, and some benchmarks (Dhrystone) show an even bigger improvement.
My suspicion (which should be taken as an educated guess) is that 0xbench's 3D benchmark ends up doing things on the CPU that it could also be doing in the GPU, including fairly heavy use of memcpy() and friends (given those are what we optimized most).
So why did we run a demo on what is likely a bogus benchmark?
First of all, because it looks good. Lots of non-tech people there - we could have run Sunspider or Dhrystone, but even though the numbers would have been more impressive, people would just have gone by without really looking - numbers are "boring", but seeing one box running circles around another is "impressive", even though both show the exact same thing.
Second, this is a repeat of an earlier demo - in last year's Connect at Cambourne, we did an identical "racing demo" (showing our build to be around 10% faster), and at the time, there just weren't that many benchmarks available to choose from.
Third, as bogus as the benchmark may be, its result pretty much matches what we're seeing in other benchmarks - and the results are consistent, so while it doesn't test what it probably intended to test (GPU performance), it does test something (probably a mix of calculations done in software, memcpy() performance and others, with a bit of GPU performance thrown in).